2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Changes: Pedro Roque : Retransmit queue handled by TCP.
25 * : Fragmentation on mtu decrease
26 * : Segment collapse on retransmit
29 * Linus Torvalds : send_delayed_ack
30 * David S. Miller : Charge memory using the right skb
31 * during syn/ack processing.
32 * David S. Miller : Output engine completely rewritten.
33 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
34 * Cacophonix Gaul : draft-minshall-nagle-01
35 * J Hadi Salim : ECN support
41 #include <linux/compiler.h>
42 #include <linux/module.h>
43 #include <linux/smp_lock.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
53 /* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
57 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
59 int sysctl_tcp_mtu_probing __read_mostly = 0;
60 int sysctl_tcp_base_mss __read_mostly = 512;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
65 static void update_send_head(struct sock *sk, struct tcp_sock *tp,
68 tcp_advance_send_head(sk, skb);
69 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
70 tcp_packets_out_inc(sk, tp, skb);
73 /* SND.NXT, if window was not shrunk.
74 * If window has been shrunk, what should we make? It is not clear at all.
75 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
76 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
77 * invalid. OK, let's make this for now:
79 static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp)
81 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
84 return tp->snd_una+tp->snd_wnd;
87 /* Calculate mss to advertise in SYN segment.
88 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
90 * 1. It is independent of path mtu.
91 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
92 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
93 * attached devices, because some buggy hosts are confused by
95 * 4. We do not make 3, we advertise MSS, calculated from first
96 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
97 * This may be overridden via information stored in routing table.
98 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
99 * probably even Jumbo".
101 static __u16 tcp_advertise_mss(struct sock *sk)
103 struct tcp_sock *tp = tcp_sk(sk);
104 struct dst_entry *dst = __sk_dst_get(sk);
105 int mss = tp->advmss;
107 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
108 mss = dst_metric(dst, RTAX_ADVMSS);
115 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
116 * This is the first part of cwnd validation mechanism. */
117 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
119 struct tcp_sock *tp = tcp_sk(sk);
120 s32 delta = tcp_time_stamp - tp->lsndtime;
121 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
122 u32 cwnd = tp->snd_cwnd;
124 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
126 tp->snd_ssthresh = tcp_current_ssthresh(sk);
127 restart_cwnd = min(restart_cwnd, cwnd);
129 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
131 tp->snd_cwnd = max(cwnd, restart_cwnd);
132 tp->snd_cwnd_stamp = tcp_time_stamp;
133 tp->snd_cwnd_used = 0;
136 static void tcp_event_data_sent(struct tcp_sock *tp,
137 struct sk_buff *skb, struct sock *sk)
139 struct inet_connection_sock *icsk = inet_csk(sk);
140 const u32 now = tcp_time_stamp;
142 if (sysctl_tcp_slow_start_after_idle &&
143 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
144 tcp_cwnd_restart(sk, __sk_dst_get(sk));
148 /* If it is a reply for ato after last received
149 * packet, enter pingpong mode.
151 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
152 icsk->icsk_ack.pingpong = 1;
155 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
157 tcp_dec_quickack_mode(sk, pkts);
158 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
161 /* Determine a window scaling and initial window to offer.
162 * Based on the assumption that the given amount of space
163 * will be offered. Store the results in the tp structure.
164 * NOTE: for smooth operation initial space offering should
165 * be a multiple of mss if possible. We assume here that mss >= 1.
166 * This MUST be enforced by all callers.
168 void tcp_select_initial_window(int __space, __u32 mss,
169 __u32 *rcv_wnd, __u32 *window_clamp,
170 int wscale_ok, __u8 *rcv_wscale)
172 unsigned int space = (__space < 0 ? 0 : __space);
174 /* If no clamp set the clamp to the max possible scaled window */
175 if (*window_clamp == 0)
176 (*window_clamp) = (65535 << 14);
177 space = min(*window_clamp, space);
179 /* Quantize space offering to a multiple of mss if possible. */
181 space = (space / mss) * mss;
183 /* NOTE: offering an initial window larger than 32767
184 * will break some buggy TCP stacks. If the admin tells us
185 * it is likely we could be speaking with such a buggy stack
186 * we will truncate our initial window offering to 32K-1
187 * unless the remote has sent us a window scaling option,
188 * which we interpret as a sign the remote TCP is not
189 * misinterpreting the window field as a signed quantity.
191 if (sysctl_tcp_workaround_signed_windows)
192 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
198 /* Set window scaling on max possible window
199 * See RFC1323 for an explanation of the limit to 14
201 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
202 space = min_t(u32, space, *window_clamp);
203 while (space > 65535 && (*rcv_wscale) < 14) {
209 /* Set initial window to value enough for senders,
210 * following RFC2414. Senders, not following this RFC,
211 * will be satisfied with 2.
213 if (mss > (1<<*rcv_wscale)) {
219 if (*rcv_wnd > init_cwnd*mss)
220 *rcv_wnd = init_cwnd*mss;
223 /* Set the clamp no higher than max representable value */
224 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
227 /* Chose a new window to advertise, update state in tcp_sock for the
228 * socket, and return result with RFC1323 scaling applied. The return
229 * value can be stuffed directly into th->window for an outgoing
232 static u16 tcp_select_window(struct sock *sk)
234 struct tcp_sock *tp = tcp_sk(sk);
235 u32 cur_win = tcp_receive_window(tp);
236 u32 new_win = __tcp_select_window(sk);
238 /* Never shrink the offered window */
239 if(new_win < cur_win) {
240 /* Danger Will Robinson!
241 * Don't update rcv_wup/rcv_wnd here or else
242 * we will not be able to advertise a zero
243 * window in time. --DaveM
245 * Relax Will Robinson.
249 tp->rcv_wnd = new_win;
250 tp->rcv_wup = tp->rcv_nxt;
252 /* Make sure we do not exceed the maximum possible
255 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
256 new_win = min(new_win, MAX_TCP_WINDOW);
258 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
260 /* RFC1323 scaling applied */
261 new_win >>= tp->rx_opt.rcv_wscale;
263 /* If we advertise zero window, disable fast path. */
270 static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
271 __u32 tstamp, __u8 **md5_hash)
273 if (tp->rx_opt.tstamp_ok) {
274 *ptr++ = htonl((TCPOPT_NOP << 24) |
276 (TCPOPT_TIMESTAMP << 8) |
278 *ptr++ = htonl(tstamp);
279 *ptr++ = htonl(tp->rx_opt.ts_recent);
281 if (tp->rx_opt.eff_sacks) {
282 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
285 *ptr++ = htonl((TCPOPT_NOP << 24) |
288 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
289 TCPOLEN_SACK_PERBLOCK)));
290 for(this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
291 *ptr++ = htonl(sp[this_sack].start_seq);
292 *ptr++ = htonl(sp[this_sack].end_seq);
294 if (tp->rx_opt.dsack) {
295 tp->rx_opt.dsack = 0;
296 tp->rx_opt.eff_sacks--;
299 #ifdef CONFIG_TCP_MD5SIG
301 *ptr++ = htonl((TCPOPT_NOP << 24) |
303 (TCPOPT_MD5SIG << 8) |
305 *md5_hash = (__u8 *)ptr;
310 /* Construct a tcp options header for a SYN or SYN_ACK packet.
311 * If this is every changed make sure to change the definition of
312 * MAX_SYN_SIZE to match the new maximum number of options that you
315 * Note - that with the RFC2385 TCP option, we make room for the
316 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
317 * location to be filled is passed back up.
319 static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
320 int offer_wscale, int wscale, __u32 tstamp,
321 __u32 ts_recent, __u8 **md5_hash)
323 /* We always get an MSS option.
324 * The option bytes which will be seen in normal data
325 * packets should timestamps be used, must be in the MSS
326 * advertised. But we subtract them from tp->mss_cache so
327 * that calculations in tcp_sendmsg are simpler etc.
328 * So account for this fact here if necessary. If we
329 * don't do this correctly, as a receiver we won't
330 * recognize data packets as being full sized when we
331 * should, and thus we won't abide by the delayed ACK
333 * SACKs don't matter, we never delay an ACK when we
334 * have any of those going out.
336 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
339 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
340 (TCPOLEN_SACK_PERM << 16) |
341 (TCPOPT_TIMESTAMP << 8) |
344 *ptr++ = htonl((TCPOPT_NOP << 24) |
346 (TCPOPT_TIMESTAMP << 8) |
348 *ptr++ = htonl(tstamp); /* TSVAL */
349 *ptr++ = htonl(ts_recent); /* TSECR */
351 *ptr++ = htonl((TCPOPT_NOP << 24) |
353 (TCPOPT_SACK_PERM << 8) |
356 *ptr++ = htonl((TCPOPT_NOP << 24) |
357 (TCPOPT_WINDOW << 16) |
358 (TCPOLEN_WINDOW << 8) |
360 #ifdef CONFIG_TCP_MD5SIG
362 * If MD5 is enabled, then we set the option, and include the size
363 * (always 18). The actual MD5 hash is added just before the
367 *ptr++ = htonl((TCPOPT_NOP << 24) |
369 (TCPOPT_MD5SIG << 8) |
371 *md5_hash = (__u8 *) ptr;
376 /* This routine actually transmits TCP packets queued in by
377 * tcp_do_sendmsg(). This is used by both the initial
378 * transmission and possible later retransmissions.
379 * All SKB's seen here are completely headerless. It is our
380 * job to build the TCP header, and pass the packet down to
381 * IP so it can do the same plus pass the packet off to the
384 * We are working here with either a clone of the original
385 * SKB, or a fresh unique copy made by the retransmit engine.
387 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
389 const struct inet_connection_sock *icsk = inet_csk(sk);
390 struct inet_sock *inet;
392 struct tcp_skb_cb *tcb;
394 #ifdef CONFIG_TCP_MD5SIG
395 struct tcp_md5sig_key *md5;
396 __u8 *md5_hash_location;
402 BUG_ON(!skb || !tcp_skb_pcount(skb));
404 /* If congestion control is doing timestamping, we must
405 * take such a timestamp before we potentially clone/copy.
407 if (icsk->icsk_ca_ops->rtt_sample)
408 __net_timestamp(skb);
410 if (likely(clone_it)) {
411 if (unlikely(skb_cloned(skb)))
412 skb = pskb_copy(skb, gfp_mask);
414 skb = skb_clone(skb, gfp_mask);
421 tcb = TCP_SKB_CB(skb);
422 tcp_header_size = tp->tcp_header_len;
424 #define SYSCTL_FLAG_TSTAMPS 0x1
425 #define SYSCTL_FLAG_WSCALE 0x2
426 #define SYSCTL_FLAG_SACK 0x4
429 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
430 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
431 if(sysctl_tcp_timestamps) {
432 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
433 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
435 if (sysctl_tcp_window_scaling) {
436 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
437 sysctl_flags |= SYSCTL_FLAG_WSCALE;
439 if (sysctl_tcp_sack) {
440 sysctl_flags |= SYSCTL_FLAG_SACK;
441 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
442 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
444 } else if (unlikely(tp->rx_opt.eff_sacks)) {
445 /* A SACK is 2 pad bytes, a 2 byte header, plus
446 * 2 32-bit sequence numbers for each SACK block.
448 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
449 (tp->rx_opt.eff_sacks *
450 TCPOLEN_SACK_PERBLOCK));
453 if (tcp_packets_in_flight(tp) == 0)
454 tcp_ca_event(sk, CA_EVENT_TX_START);
456 #ifdef CONFIG_TCP_MD5SIG
458 * Are we doing MD5 on this segment? If so - make
461 md5 = tp->af_specific->md5_lookup(sk, sk);
463 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
466 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
468 skb_set_owner_w(skb, sk);
470 /* Build TCP header and checksum it. */
471 th->source = inet->sport;
472 th->dest = inet->dport;
473 th->seq = htonl(tcb->seq);
474 th->ack_seq = htonl(tp->rcv_nxt);
475 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
478 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
479 /* RFC1323: The window in SYN & SYN/ACK segments
482 th->window = htons(min(tp->rcv_wnd, 65535U));
484 th->window = htons(tcp_select_window(sk));
489 if (unlikely(tp->urg_mode &&
490 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
491 th->urg_ptr = htons(tp->snd_up-tcb->seq);
495 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
496 tcp_syn_build_options((__be32 *)(th + 1),
497 tcp_advertise_mss(sk),
498 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
499 (sysctl_flags & SYSCTL_FLAG_SACK),
500 (sysctl_flags & SYSCTL_FLAG_WSCALE),
501 tp->rx_opt.rcv_wscale,
503 tp->rx_opt.ts_recent,
505 #ifdef CONFIG_TCP_MD5SIG
506 md5 ? &md5_hash_location :
510 tcp_build_and_update_options((__be32 *)(th + 1),
512 #ifdef CONFIG_TCP_MD5SIG
513 md5 ? &md5_hash_location :
516 TCP_ECN_send(sk, tp, skb, tcp_header_size);
519 #ifdef CONFIG_TCP_MD5SIG
520 /* Calculate the MD5 hash, as we have all we need now */
522 tp->af_specific->calc_md5_hash(md5_hash_location,
531 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
533 if (likely(tcb->flags & TCPCB_FLAG_ACK))
534 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
536 if (skb->len != tcp_header_size)
537 tcp_event_data_sent(tp, skb, sk);
539 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
540 TCP_INC_STATS(TCP_MIB_OUTSEGS);
542 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
543 if (likely(err <= 0))
546 tcp_enter_cwr(sk, 1);
548 return net_xmit_eval(err);
550 #undef SYSCTL_FLAG_TSTAMPS
551 #undef SYSCTL_FLAG_WSCALE
552 #undef SYSCTL_FLAG_SACK
556 /* This routine just queue's the buffer
558 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
559 * otherwise socket can stall.
561 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
563 struct tcp_sock *tp = tcp_sk(sk);
565 /* Advance write_seq and place onto the write_queue. */
566 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
567 skb_header_release(skb);
568 tcp_add_write_queue_tail(sk, skb);
569 sk_charge_skb(sk, skb);
572 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
574 if (skb->len <= mss_now || !sk_can_gso(sk)) {
575 /* Avoid the costly divide in the normal
578 skb_shinfo(skb)->gso_segs = 1;
579 skb_shinfo(skb)->gso_size = 0;
580 skb_shinfo(skb)->gso_type = 0;
584 factor = skb->len + (mss_now - 1);
586 skb_shinfo(skb)->gso_segs = factor;
587 skb_shinfo(skb)->gso_size = mss_now;
588 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
592 /* Function to create two new TCP segments. Shrinks the given segment
593 * to the specified size and appends a new segment with the rest of the
594 * packet to the list. This won't be called frequently, I hope.
595 * Remember, these are still headerless SKBs at this point.
597 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
599 struct tcp_sock *tp = tcp_sk(sk);
600 struct sk_buff *buff;
601 int nsize, old_factor;
605 BUG_ON(len > skb->len);
607 clear_all_retrans_hints(tp);
608 nsize = skb_headlen(skb) - len;
612 if (skb_cloned(skb) &&
613 skb_is_nonlinear(skb) &&
614 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
617 /* Get a new skb... force flag on. */
618 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
620 return -ENOMEM; /* We'll just try again later. */
622 sk_charge_skb(sk, buff);
623 nlen = skb->len - len - nsize;
624 buff->truesize += nlen;
625 skb->truesize -= nlen;
627 /* Correct the sequence numbers. */
628 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
629 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
630 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
632 /* PSH and FIN should only be set in the second packet. */
633 flags = TCP_SKB_CB(skb)->flags;
634 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
635 TCP_SKB_CB(buff)->flags = flags;
636 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
637 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
639 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
640 /* Copy and checksum data tail into the new buffer. */
641 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
646 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
648 skb->ip_summed = CHECKSUM_PARTIAL;
649 skb_split(skb, buff, len);
652 buff->ip_summed = skb->ip_summed;
654 /* Looks stupid, but our code really uses when of
655 * skbs, which it never sent before. --ANK
657 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
658 buff->tstamp = skb->tstamp;
660 old_factor = tcp_skb_pcount(skb);
662 /* Fix up tso_factor for both original and new SKB. */
663 tcp_set_skb_tso_segs(sk, skb, mss_now);
664 tcp_set_skb_tso_segs(sk, buff, mss_now);
666 /* If this packet has been sent out already, we must
667 * adjust the various packet counters.
669 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
670 int diff = old_factor - tcp_skb_pcount(skb) -
671 tcp_skb_pcount(buff);
673 tp->packets_out -= diff;
675 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
676 tp->sacked_out -= diff;
677 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
678 tp->retrans_out -= diff;
680 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
681 tp->lost_out -= diff;
682 tp->left_out -= diff;
686 /* Adjust Reno SACK estimate. */
687 if (!tp->rx_opt.sack_ok) {
688 tp->sacked_out -= diff;
689 if ((int)tp->sacked_out < 0)
691 tcp_sync_left_out(tp);
694 tp->fackets_out -= diff;
695 if ((int)tp->fackets_out < 0)
700 /* Link BUFF into the send queue. */
701 skb_header_release(buff);
702 tcp_insert_write_queue_after(skb, buff, sk);
707 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
708 * eventually). The difference is that pulled data not copied, but
709 * immediately discarded.
711 static void __pskb_trim_head(struct sk_buff *skb, int len)
717 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
718 if (skb_shinfo(skb)->frags[i].size <= eat) {
719 put_page(skb_shinfo(skb)->frags[i].page);
720 eat -= skb_shinfo(skb)->frags[i].size;
722 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
724 skb_shinfo(skb)->frags[k].page_offset += eat;
725 skb_shinfo(skb)->frags[k].size -= eat;
731 skb_shinfo(skb)->nr_frags = k;
733 skb->tail = skb->data;
734 skb->data_len -= len;
735 skb->len = skb->data_len;
738 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
740 if (skb_cloned(skb) &&
741 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
744 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
745 if (unlikely(len < skb_headlen(skb)))
746 __skb_pull(skb, len);
748 __pskb_trim_head(skb, len - skb_headlen(skb));
750 TCP_SKB_CB(skb)->seq += len;
751 skb->ip_summed = CHECKSUM_PARTIAL;
753 skb->truesize -= len;
754 sk->sk_wmem_queued -= len;
755 sk->sk_forward_alloc += len;
756 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
758 /* Any change of skb->len requires recalculation of tso
761 if (tcp_skb_pcount(skb) > 1)
762 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
767 /* Not accounting for SACKs here. */
768 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
770 struct tcp_sock *tp = tcp_sk(sk);
771 struct inet_connection_sock *icsk = inet_csk(sk);
774 /* Calculate base mss without TCP options:
775 It is MMS_S - sizeof(tcphdr) of rfc1122
777 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
779 /* Clamp it (mss_clamp does not include tcp options) */
780 if (mss_now > tp->rx_opt.mss_clamp)
781 mss_now = tp->rx_opt.mss_clamp;
783 /* Now subtract optional transport overhead */
784 mss_now -= icsk->icsk_ext_hdr_len;
786 /* Then reserve room for full set of TCP options and 8 bytes of data */
790 /* Now subtract TCP options size, not including SACKs */
791 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
796 /* Inverse of above */
797 int tcp_mss_to_mtu(struct sock *sk, int mss)
799 struct tcp_sock *tp = tcp_sk(sk);
800 struct inet_connection_sock *icsk = inet_csk(sk);
805 icsk->icsk_ext_hdr_len +
806 icsk->icsk_af_ops->net_header_len;
811 void tcp_mtup_init(struct sock *sk)
813 struct tcp_sock *tp = tcp_sk(sk);
814 struct inet_connection_sock *icsk = inet_csk(sk);
816 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
817 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
818 icsk->icsk_af_ops->net_header_len;
819 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
820 icsk->icsk_mtup.probe_size = 0;
823 /* This function synchronize snd mss to current pmtu/exthdr set.
825 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
826 for TCP options, but includes only bare TCP header.
828 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
829 It is minimum of user_mss and mss received with SYN.
830 It also does not include TCP options.
832 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
834 tp->mss_cache is current effective sending mss, including
835 all tcp options except for SACKs. It is evaluated,
836 taking into account current pmtu, but never exceeds
837 tp->rx_opt.mss_clamp.
839 NOTE1. rfc1122 clearly states that advertised MSS
840 DOES NOT include either tcp or ip options.
842 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
843 are READ ONLY outside this function. --ANK (980731)
846 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
848 struct tcp_sock *tp = tcp_sk(sk);
849 struct inet_connection_sock *icsk = inet_csk(sk);
852 if (icsk->icsk_mtup.search_high > pmtu)
853 icsk->icsk_mtup.search_high = pmtu;
855 mss_now = tcp_mtu_to_mss(sk, pmtu);
857 /* Bound mss with half of window */
858 if (tp->max_window && mss_now > (tp->max_window>>1))
859 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
861 /* And store cached results */
862 icsk->icsk_pmtu_cookie = pmtu;
863 if (icsk->icsk_mtup.enabled)
864 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
865 tp->mss_cache = mss_now;
870 /* Compute the current effective MSS, taking SACKs and IP options,
871 * and even PMTU discovery events into account.
873 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
874 * cannot be large. However, taking into account rare use of URG, this
877 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
879 struct tcp_sock *tp = tcp_sk(sk);
880 struct dst_entry *dst = __sk_dst_get(sk);
885 mss_now = tp->mss_cache;
887 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
891 u32 mtu = dst_mtu(dst);
892 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
893 mss_now = tcp_sync_mss(sk, mtu);
896 if (tp->rx_opt.eff_sacks)
897 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
898 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
900 #ifdef CONFIG_TCP_MD5SIG
901 if (tp->af_specific->md5_lookup(sk, sk))
902 mss_now -= TCPOLEN_MD5SIG_ALIGNED;
905 xmit_size_goal = mss_now;
908 xmit_size_goal = (65535 -
909 inet_csk(sk)->icsk_af_ops->net_header_len -
910 inet_csk(sk)->icsk_ext_hdr_len -
913 if (tp->max_window &&
914 (xmit_size_goal > (tp->max_window >> 1)))
915 xmit_size_goal = max((tp->max_window >> 1),
916 68U - tp->tcp_header_len);
918 xmit_size_goal -= (xmit_size_goal % mss_now);
920 tp->xmit_size_goal = xmit_size_goal;
925 /* Congestion window validation. (RFC2861) */
927 static void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
929 __u32 packets_out = tp->packets_out;
931 if (packets_out >= tp->snd_cwnd) {
932 /* Network is feed fully. */
933 tp->snd_cwnd_used = 0;
934 tp->snd_cwnd_stamp = tcp_time_stamp;
936 /* Network starves. */
937 if (tp->packets_out > tp->snd_cwnd_used)
938 tp->snd_cwnd_used = tp->packets_out;
940 if (sysctl_tcp_slow_start_after_idle &&
941 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
942 tcp_cwnd_application_limited(sk);
946 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
948 u32 window, cwnd_len;
950 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
951 cwnd_len = mss_now * cwnd;
952 return min(window, cwnd_len);
955 /* Can at least one segment of SKB be sent right now, according to the
956 * congestion window rules? If so, return how many segments are allowed.
958 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
962 /* Don't be strict about the congestion window for the final FIN. */
963 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
964 tcp_skb_pcount(skb) == 1)
967 in_flight = tcp_packets_in_flight(tp);
969 if (in_flight < cwnd)
970 return (cwnd - in_flight);
975 /* This must be invoked the first time we consider transmitting
978 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
980 int tso_segs = tcp_skb_pcount(skb);
984 tcp_skb_mss(skb) != mss_now)) {
985 tcp_set_skb_tso_segs(sk, skb, mss_now);
986 tso_segs = tcp_skb_pcount(skb);
991 static inline int tcp_minshall_check(const struct tcp_sock *tp)
993 return after(tp->snd_sml,tp->snd_una) &&
994 !after(tp->snd_sml, tp->snd_nxt);
997 /* Return 0, if packet can be sent now without violation Nagle's rules:
998 * 1. It is full sized.
999 * 2. Or it contains FIN. (already checked by caller)
1000 * 3. Or TCP_NODELAY was set.
1001 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1002 * With Minshall's modification: all sent small packets are ACKed.
1005 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1006 const struct sk_buff *skb,
1007 unsigned mss_now, int nonagle)
1009 return (skb->len < mss_now &&
1010 ((nonagle&TCP_NAGLE_CORK) ||
1013 tcp_minshall_check(tp))));
1016 /* Return non-zero if the Nagle test allows this packet to be
1019 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1020 unsigned int cur_mss, int nonagle)
1022 /* Nagle rule does not apply to frames, which sit in the middle of the
1023 * write_queue (they have no chances to get new data).
1025 * This is implemented in the callers, where they modify the 'nonagle'
1026 * argument based upon the location of SKB in the send queue.
1028 if (nonagle & TCP_NAGLE_PUSH)
1031 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1033 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1036 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1042 /* Does at least the first segment of SKB fit into the send window? */
1043 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
1045 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1047 if (skb->len > cur_mss)
1048 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1050 return !after(end_seq, tp->snd_una + tp->snd_wnd);
1053 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1054 * should be put on the wire right now. If so, it returns the number of
1055 * packets allowed by the congestion window.
1057 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1058 unsigned int cur_mss, int nonagle)
1060 struct tcp_sock *tp = tcp_sk(sk);
1061 unsigned int cwnd_quota;
1063 tcp_init_tso_segs(sk, skb, cur_mss);
1065 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1068 cwnd_quota = tcp_cwnd_test(tp, skb);
1070 !tcp_snd_wnd_test(tp, skb, cur_mss))
1076 int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
1078 struct sk_buff *skb = tcp_send_head(sk);
1081 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1082 (tcp_skb_is_last(sk, skb) ?
1087 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1088 * which is put after SKB on the list. It is very much like
1089 * tcp_fragment() except that it may make several kinds of assumptions
1090 * in order to speed up the splitting operation. In particular, we
1091 * know that all the data is in scatter-gather pages, and that the
1092 * packet has never been sent out before (and thus is not cloned).
1094 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
1096 struct sk_buff *buff;
1097 int nlen = skb->len - len;
1100 /* All of a TSO frame must be composed of paged data. */
1101 if (skb->len != skb->data_len)
1102 return tcp_fragment(sk, skb, len, mss_now);
1104 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
1105 if (unlikely(buff == NULL))
1108 sk_charge_skb(sk, buff);
1109 buff->truesize += nlen;
1110 skb->truesize -= nlen;
1112 /* Correct the sequence numbers. */
1113 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1114 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1115 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1117 /* PSH and FIN should only be set in the second packet. */
1118 flags = TCP_SKB_CB(skb)->flags;
1119 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1120 TCP_SKB_CB(buff)->flags = flags;
1122 /* This packet was never sent out yet, so no SACK bits. */
1123 TCP_SKB_CB(buff)->sacked = 0;
1125 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1126 skb_split(skb, buff, len);
1128 /* Fix up tso_factor for both original and new SKB. */
1129 tcp_set_skb_tso_segs(sk, skb, mss_now);
1130 tcp_set_skb_tso_segs(sk, buff, mss_now);
1132 /* Link BUFF into the send queue. */
1133 skb_header_release(buff);
1134 tcp_insert_write_queue_after(skb, buff, sk);
1139 /* Try to defer sending, if possible, in order to minimize the amount
1140 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1142 * This algorithm is from John Heffner.
1144 static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
1146 const struct inet_connection_sock *icsk = inet_csk(sk);
1147 u32 send_win, cong_win, limit, in_flight;
1149 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1152 if (icsk->icsk_ca_state != TCP_CA_Open)
1155 /* Defer for less than two clock ticks. */
1156 if (!tp->tso_deferred && ((jiffies<<1)>>1) - (tp->tso_deferred>>1) > 1)
1159 in_flight = tcp_packets_in_flight(tp);
1161 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1162 (tp->snd_cwnd <= in_flight));
1164 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1166 /* From in_flight test above, we know that cwnd > in_flight. */
1167 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1169 limit = min(send_win, cong_win);
1171 /* If a full-sized TSO skb can be sent, do it. */
1175 if (sysctl_tcp_tso_win_divisor) {
1176 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1178 /* If at least some fraction of a window is available,
1181 chunk /= sysctl_tcp_tso_win_divisor;
1185 /* Different approach, try not to defer past a single
1186 * ACK. Receiver should ACK every other full sized
1187 * frame, so if we have space for more than 3 frames
1190 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1194 /* Ok, it looks like it is advisable to defer. */
1195 tp->tso_deferred = 1 | (jiffies<<1);
1200 tp->tso_deferred = 0;
1204 /* Create a new MTU probe if we are ready.
1205 * Returns 0 if we should wait to probe (no cwnd available),
1206 * 1 if a probe was sent,
1208 static int tcp_mtu_probe(struct sock *sk)
1210 struct tcp_sock *tp = tcp_sk(sk);
1211 struct inet_connection_sock *icsk = inet_csk(sk);
1212 struct sk_buff *skb, *nskb, *next;
1219 /* Not currently probing/verifying,
1221 * have enough cwnd, and
1222 * not SACKing (the variable headers throw things off) */
1223 if (!icsk->icsk_mtup.enabled ||
1224 icsk->icsk_mtup.probe_size ||
1225 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1226 tp->snd_cwnd < 11 ||
1227 tp->rx_opt.eff_sacks)
1230 /* Very simple search strategy: just double the MSS. */
1231 mss_now = tcp_current_mss(sk, 0);
1232 probe_size = 2*tp->mss_cache;
1233 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1234 /* TODO: set timer for probe_converge_event */
1238 /* Have enough data in the send queue to probe? */
1240 if ((skb = tcp_send_head(sk)) == NULL)
1242 while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb))
1243 skb = tcp_write_queue_next(sk, skb);
1244 if (len < probe_size)
1247 /* Receive window check. */
1248 if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) {
1249 if (tp->snd_wnd < probe_size)
1255 /* Do we need to wait to drain cwnd? */
1256 pif = tcp_packets_in_flight(tp);
1257 if (pif + 2 > tp->snd_cwnd) {
1258 /* With no packets in flight, don't stall. */
1265 /* We're allowed to probe. Build it now. */
1266 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1268 sk_charge_skb(sk, nskb);
1270 skb = tcp_send_head(sk);
1271 tcp_insert_write_queue_before(nskb, skb, sk);
1272 tcp_advance_send_head(sk, skb);
1274 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1275 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1276 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1277 TCP_SKB_CB(nskb)->sacked = 0;
1279 nskb->ip_summed = skb->ip_summed;
1282 while (len < probe_size) {
1283 next = tcp_write_queue_next(sk, skb);
1285 copy = min_t(int, skb->len, probe_size - len);
1286 if (nskb->ip_summed)
1287 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1289 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1290 skb_put(nskb, copy), copy, nskb->csum);
1292 if (skb->len <= copy) {
1293 /* We've eaten all the data from this skb.
1295 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1296 tcp_unlink_write_queue(skb, sk);
1297 sk_stream_free_skb(sk, skb);
1299 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1300 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1301 if (!skb_shinfo(skb)->nr_frags) {
1302 skb_pull(skb, copy);
1303 if (skb->ip_summed != CHECKSUM_PARTIAL)
1304 skb->csum = csum_partial(skb->data, skb->len, 0);
1306 __pskb_trim_head(skb, copy);
1307 tcp_set_skb_tso_segs(sk, skb, mss_now);
1309 TCP_SKB_CB(skb)->seq += copy;
1315 tcp_init_tso_segs(sk, nskb, nskb->len);
1317 /* We're ready to send. If this fails, the probe will
1318 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1319 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1320 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1321 /* Decrement cwnd here because we are sending
1322 * effectively two packets. */
1324 update_send_head(sk, tp, nskb);
1326 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1327 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1328 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1337 /* This routine writes packets to the network. It advances the
1338 * send_head. This happens as incoming acks open up the remote
1341 * Returns 1, if no segments are in flight and we have queued segments, but
1342 * cannot send anything now because of SWS or another problem.
1344 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1346 struct tcp_sock *tp = tcp_sk(sk);
1347 struct sk_buff *skb;
1348 unsigned int tso_segs, sent_pkts;
1352 /* If we are closed, the bytes will have to remain here.
1353 * In time closedown will finish, we empty the write queue and all
1356 if (unlikely(sk->sk_state == TCP_CLOSE))
1361 /* Do MTU probing. */
1362 if ((result = tcp_mtu_probe(sk)) == 0) {
1364 } else if (result > 0) {
1368 while ((skb = tcp_send_head(sk))) {
1371 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1374 cwnd_quota = tcp_cwnd_test(tp, skb);
1378 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1381 if (tso_segs == 1) {
1382 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1383 (tcp_skb_is_last(sk, skb) ?
1384 nonagle : TCP_NAGLE_PUSH))))
1387 if (tcp_tso_should_defer(sk, tp, skb))
1393 limit = tcp_window_allows(tp, skb,
1394 mss_now, cwnd_quota);
1396 if (skb->len < limit) {
1397 unsigned int trim = skb->len % mss_now;
1400 limit = skb->len - trim;
1404 if (skb->len > limit &&
1405 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1408 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1410 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1413 /* Advance the send_head. This one is sent out.
1414 * This call will increment packets_out.
1416 update_send_head(sk, tp, skb);
1418 tcp_minshall_update(tp, mss_now, skb);
1422 if (likely(sent_pkts)) {
1423 tcp_cwnd_validate(sk, tp);
1426 return !tp->packets_out && tcp_send_head(sk);
1429 /* Push out any pending frames which were held back due to
1430 * TCP_CORK or attempt at coalescing tiny packets.
1431 * The socket must be locked by the caller.
1433 void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
1434 unsigned int cur_mss, int nonagle)
1436 struct sk_buff *skb = tcp_send_head(sk);
1439 if (tcp_write_xmit(sk, cur_mss, nonagle))
1440 tcp_check_probe_timer(sk, tp);
1444 /* Send _single_ skb sitting at the send head. This function requires
1445 * true push pending frames to setup probe timer etc.
1447 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1449 struct tcp_sock *tp = tcp_sk(sk);
1450 struct sk_buff *skb = tcp_send_head(sk);
1451 unsigned int tso_segs, cwnd_quota;
1453 BUG_ON(!skb || skb->len < mss_now);
1455 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1456 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1458 if (likely(cwnd_quota)) {
1465 limit = tcp_window_allows(tp, skb,
1466 mss_now, cwnd_quota);
1468 if (skb->len < limit) {
1469 unsigned int trim = skb->len % mss_now;
1472 limit = skb->len - trim;
1476 if (skb->len > limit &&
1477 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1480 /* Send it out now. */
1481 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1483 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1484 update_send_head(sk, tp, skb);
1485 tcp_cwnd_validate(sk, tp);
1491 /* This function returns the amount that we can raise the
1492 * usable window based on the following constraints
1494 * 1. The window can never be shrunk once it is offered (RFC 793)
1495 * 2. We limit memory per socket
1498 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1499 * RECV.NEXT + RCV.WIN fixed until:
1500 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1502 * i.e. don't raise the right edge of the window until you can raise
1503 * it at least MSS bytes.
1505 * Unfortunately, the recommended algorithm breaks header prediction,
1506 * since header prediction assumes th->window stays fixed.
1508 * Strictly speaking, keeping th->window fixed violates the receiver
1509 * side SWS prevention criteria. The problem is that under this rule
1510 * a stream of single byte packets will cause the right side of the
1511 * window to always advance by a single byte.
1513 * Of course, if the sender implements sender side SWS prevention
1514 * then this will not be a problem.
1516 * BSD seems to make the following compromise:
1518 * If the free space is less than the 1/4 of the maximum
1519 * space available and the free space is less than 1/2 mss,
1520 * then set the window to 0.
1521 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1522 * Otherwise, just prevent the window from shrinking
1523 * and from being larger than the largest representable value.
1525 * This prevents incremental opening of the window in the regime
1526 * where TCP is limited by the speed of the reader side taking
1527 * data out of the TCP receive queue. It does nothing about
1528 * those cases where the window is constrained on the sender side
1529 * because the pipeline is full.
1531 * BSD also seems to "accidentally" limit itself to windows that are a
1532 * multiple of MSS, at least until the free space gets quite small.
1533 * This would appear to be a side effect of the mbuf implementation.
1534 * Combining these two algorithms results in the observed behavior
1535 * of having a fixed window size at almost all times.
1537 * Below we obtain similar behavior by forcing the offered window to
1538 * a multiple of the mss when it is feasible to do so.
1540 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1541 * Regular options like TIMESTAMP are taken into account.
1543 u32 __tcp_select_window(struct sock *sk)
1545 struct inet_connection_sock *icsk = inet_csk(sk);
1546 struct tcp_sock *tp = tcp_sk(sk);
1547 /* MSS for the peer's data. Previous versions used mss_clamp
1548 * here. I don't know if the value based on our guesses
1549 * of peer's MSS is better for the performance. It's more correct
1550 * but may be worse for the performance because of rcv_mss
1551 * fluctuations. --SAW 1998/11/1
1553 int mss = icsk->icsk_ack.rcv_mss;
1554 int free_space = tcp_space(sk);
1555 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1558 if (mss > full_space)
1561 if (free_space < full_space/2) {
1562 icsk->icsk_ack.quick = 0;
1564 if (tcp_memory_pressure)
1565 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1567 if (free_space < mss)
1571 if (free_space > tp->rcv_ssthresh)
1572 free_space = tp->rcv_ssthresh;
1574 /* Don't do rounding if we are using window scaling, since the
1575 * scaled window will not line up with the MSS boundary anyway.
1577 window = tp->rcv_wnd;
1578 if (tp->rx_opt.rcv_wscale) {
1579 window = free_space;
1581 /* Advertise enough space so that it won't get scaled away.
1582 * Import case: prevent zero window announcement if
1583 * 1<<rcv_wscale > mss.
1585 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1586 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1587 << tp->rx_opt.rcv_wscale);
1589 /* Get the largest window that is a nice multiple of mss.
1590 * Window clamp already applied above.
1591 * If our current window offering is within 1 mss of the
1592 * free space we just keep it. This prevents the divide
1593 * and multiply from happening most of the time.
1594 * We also don't do any window rounding when the free space
1597 if (window <= free_space - mss || window > free_space)
1598 window = (free_space/mss)*mss;
1599 else if (mss == full_space &&
1600 free_space > window + full_space/2)
1601 window = free_space;
1607 /* Attempt to collapse two adjacent SKB's during retransmission. */
1608 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1610 struct tcp_sock *tp = tcp_sk(sk);
1611 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1613 /* The first test we must make is that neither of these two
1614 * SKB's are still referenced by someone else.
1616 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1617 int skb_size = skb->len, next_skb_size = next_skb->len;
1618 u16 flags = TCP_SKB_CB(skb)->flags;
1620 /* Also punt if next skb has been SACK'd. */
1621 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1624 /* Next skb is out of window. */
1625 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1628 /* Punt if not enough space exists in the first SKB for
1629 * the data in the second, or the total combined payload
1630 * would exceed the MSS.
1632 if ((next_skb_size > skb_tailroom(skb)) ||
1633 ((skb_size + next_skb_size) > mss_now))
1636 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1637 tcp_skb_pcount(next_skb) != 1);
1639 /* changing transmit queue under us so clear hints */
1640 clear_all_retrans_hints(tp);
1642 /* Ok. We will be able to collapse the packet. */
1643 tcp_unlink_write_queue(next_skb, sk);
1645 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1647 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1648 skb->ip_summed = CHECKSUM_PARTIAL;
1650 if (skb->ip_summed != CHECKSUM_PARTIAL)
1651 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1653 /* Update sequence range on original skb. */
1654 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1656 /* Merge over control information. */
1657 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1658 TCP_SKB_CB(skb)->flags = flags;
1660 /* All done, get rid of second SKB and account for it so
1661 * packet counting does not break.
1663 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1664 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1665 tp->retrans_out -= tcp_skb_pcount(next_skb);
1666 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1667 tp->lost_out -= tcp_skb_pcount(next_skb);
1668 tp->left_out -= tcp_skb_pcount(next_skb);
1670 /* Reno case is special. Sigh... */
1671 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1672 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1673 tp->left_out -= tcp_skb_pcount(next_skb);
1676 /* Not quite right: it can be > snd.fack, but
1677 * it is better to underestimate fackets.
1679 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1680 tcp_packets_out_dec(tp, next_skb);
1681 sk_stream_free_skb(sk, next_skb);
1685 /* Do a simple retransmit without using the backoff mechanisms in
1686 * tcp_timer. This is used for path mtu discovery.
1687 * The socket is already locked here.
1689 void tcp_simple_retransmit(struct sock *sk)
1691 const struct inet_connection_sock *icsk = inet_csk(sk);
1692 struct tcp_sock *tp = tcp_sk(sk);
1693 struct sk_buff *skb;
1694 unsigned int mss = tcp_current_mss(sk, 0);
1697 tcp_for_write_queue(skb, sk) {
1698 if (skb == tcp_send_head(sk))
1700 if (skb->len > mss &&
1701 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1702 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1703 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1704 tp->retrans_out -= tcp_skb_pcount(skb);
1706 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1707 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1708 tp->lost_out += tcp_skb_pcount(skb);
1714 clear_all_retrans_hints(tp);
1719 tcp_sync_left_out(tp);
1721 /* Don't muck with the congestion window here.
1722 * Reason is that we do not increase amount of _data_
1723 * in network, but units changed and effective
1724 * cwnd/ssthresh really reduced now.
1726 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1727 tp->high_seq = tp->snd_nxt;
1728 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1729 tp->prior_ssthresh = 0;
1730 tp->undo_marker = 0;
1731 tcp_set_ca_state(sk, TCP_CA_Loss);
1733 tcp_xmit_retransmit_queue(sk);
1736 /* This retransmits one SKB. Policy decisions and retransmit queue
1737 * state updates are done by the caller. Returns non-zero if an
1738 * error occurred which prevented the send.
1740 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1742 struct tcp_sock *tp = tcp_sk(sk);
1743 struct inet_connection_sock *icsk = inet_csk(sk);
1744 unsigned int cur_mss = tcp_current_mss(sk, 0);
1747 /* Inconslusive MTU probe */
1748 if (icsk->icsk_mtup.probe_size) {
1749 icsk->icsk_mtup.probe_size = 0;
1752 /* Do not sent more than we queued. 1/4 is reserved for possible
1753 * copying overhead: fragmentation, tunneling, mangling etc.
1755 if (atomic_read(&sk->sk_wmem_alloc) >
1756 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1759 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1760 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1762 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1766 /* If receiver has shrunk his window, and skb is out of
1767 * new window, do not retransmit it. The exception is the
1768 * case, when window is shrunk to zero. In this case
1769 * our retransmit serves as a zero window probe.
1771 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1772 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1775 if (skb->len > cur_mss) {
1776 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1777 return -ENOMEM; /* We'll try again later. */
1780 /* Collapse two adjacent packets if worthwhile and we can. */
1781 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1782 (skb->len < (cur_mss >> 1)) &&
1783 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1784 (!tcp_skb_is_last(sk, skb)) &&
1785 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1786 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1787 (sysctl_tcp_retrans_collapse != 0))
1788 tcp_retrans_try_collapse(sk, skb, cur_mss);
1790 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1791 return -EHOSTUNREACH; /* Routing failure or similar. */
1793 /* Some Solaris stacks overoptimize and ignore the FIN on a
1794 * retransmit when old data is attached. So strip it off
1795 * since it is cheap to do so and saves bytes on the network.
1798 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1799 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1800 if (!pskb_trim(skb, 0)) {
1801 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1802 skb_shinfo(skb)->gso_segs = 1;
1803 skb_shinfo(skb)->gso_size = 0;
1804 skb_shinfo(skb)->gso_type = 0;
1805 skb->ip_summed = CHECKSUM_NONE;
1810 /* Make a copy, if the first transmission SKB clone we made
1811 * is still in somebody's hands, else make a clone.
1813 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1815 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1818 /* Update global TCP statistics. */
1819 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1821 tp->total_retrans++;
1823 #if FASTRETRANS_DEBUG > 0
1824 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1825 if (net_ratelimit())
1826 printk(KERN_DEBUG "retrans_out leaked.\n");
1829 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1830 tp->retrans_out += tcp_skb_pcount(skb);
1832 /* Save stamp of the first retransmit. */
1833 if (!tp->retrans_stamp)
1834 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1838 /* snd_nxt is stored to detect loss of retransmitted segment,
1839 * see tcp_input.c tcp_sacktag_write_queue().
1841 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1846 /* This gets called after a retransmit timeout, and the initially
1847 * retransmitted data is acknowledged. It tries to continue
1848 * resending the rest of the retransmit queue, until either
1849 * we've sent it all or the congestion window limit is reached.
1850 * If doing SACK, the first ACK which comes back for a timeout
1851 * based retransmit packet might feed us FACK information again.
1852 * If so, we use it to avoid unnecessarily retransmissions.
1854 void tcp_xmit_retransmit_queue(struct sock *sk)
1856 const struct inet_connection_sock *icsk = inet_csk(sk);
1857 struct tcp_sock *tp = tcp_sk(sk);
1858 struct sk_buff *skb;
1861 if (tp->retransmit_skb_hint) {
1862 skb = tp->retransmit_skb_hint;
1863 packet_cnt = tp->retransmit_cnt_hint;
1865 skb = tcp_write_queue_head(sk);
1869 /* First pass: retransmit lost packets. */
1871 tcp_for_write_queue_from(skb, sk) {
1872 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1874 if (skb == tcp_send_head(sk))
1876 /* we could do better than to assign each time */
1877 tp->retransmit_skb_hint = skb;
1878 tp->retransmit_cnt_hint = packet_cnt;
1880 /* Assume this retransmit will generate
1881 * only one packet for congestion window
1882 * calculation purposes. This works because
1883 * tcp_retransmit_skb() will chop up the
1884 * packet to be MSS sized and all the
1885 * packet counting works out.
1887 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1890 if (sacked & TCPCB_LOST) {
1891 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1892 if (tcp_retransmit_skb(sk, skb)) {
1893 tp->retransmit_skb_hint = NULL;
1896 if (icsk->icsk_ca_state != TCP_CA_Loss)
1897 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1899 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1901 if (skb == tcp_write_queue_head(sk))
1902 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1903 inet_csk(sk)->icsk_rto,
1907 packet_cnt += tcp_skb_pcount(skb);
1908 if (packet_cnt >= tp->lost_out)
1914 /* OK, demanded retransmission is finished. */
1916 /* Forward retransmissions are possible only during Recovery. */
1917 if (icsk->icsk_ca_state != TCP_CA_Recovery)
1920 /* No forward retransmissions in Reno are possible. */
1921 if (!tp->rx_opt.sack_ok)
1924 /* Yeah, we have to make difficult choice between forward transmission
1925 * and retransmission... Both ways have their merits...
1927 * For now we do not retransmit anything, while we have some new
1931 if (tcp_may_send_now(sk, tp))
1934 if (tp->forward_skb_hint) {
1935 skb = tp->forward_skb_hint;
1936 packet_cnt = tp->forward_cnt_hint;
1938 skb = tcp_write_queue_head(sk);
1942 tcp_for_write_queue_from(skb, sk) {
1943 if (skb == tcp_send_head(sk))
1945 tp->forward_cnt_hint = packet_cnt;
1946 tp->forward_skb_hint = skb;
1948 /* Similar to the retransmit loop above we
1949 * can pretend that the retransmitted SKB
1950 * we send out here will be composed of one
1951 * real MSS sized packet because tcp_retransmit_skb()
1952 * will fragment it if necessary.
1954 if (++packet_cnt > tp->fackets_out)
1957 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1960 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1963 /* Ok, retransmit it. */
1964 if (tcp_retransmit_skb(sk, skb)) {
1965 tp->forward_skb_hint = NULL;
1969 if (skb == tcp_write_queue_head(sk))
1970 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1971 inet_csk(sk)->icsk_rto,
1974 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1979 /* Send a fin. The caller locks the socket for us. This cannot be
1980 * allowed to fail queueing a FIN frame under any circumstances.
1982 void tcp_send_fin(struct sock *sk)
1984 struct tcp_sock *tp = tcp_sk(sk);
1985 struct sk_buff *skb = tcp_write_queue_tail(sk);
1988 /* Optimization, tack on the FIN if we have a queue of
1989 * unsent frames. But be careful about outgoing SACKS
1992 mss_now = tcp_current_mss(sk, 1);
1994 if (tcp_send_head(sk) != NULL) {
1995 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1996 TCP_SKB_CB(skb)->end_seq++;
1999 /* Socket is locked, keep trying until memory is available. */
2001 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
2007 /* Reserve space for headers and prepare control bits. */
2008 skb_reserve(skb, MAX_TCP_HEADER);
2010 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2011 TCP_SKB_CB(skb)->sacked = 0;
2012 skb_shinfo(skb)->gso_segs = 1;
2013 skb_shinfo(skb)->gso_size = 0;
2014 skb_shinfo(skb)->gso_type = 0;
2016 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2017 TCP_SKB_CB(skb)->seq = tp->write_seq;
2018 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2019 tcp_queue_skb(sk, skb);
2021 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
2024 /* We get here when a process closes a file descriptor (either due to
2025 * an explicit close() or as a byproduct of exit()'ing) and there
2026 * was unread data in the receive queue. This behavior is recommended
2027 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
2029 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2031 struct tcp_sock *tp = tcp_sk(sk);
2032 struct sk_buff *skb;
2034 /* NOTE: No TCP options attached and we never retransmit this. */
2035 skb = alloc_skb(MAX_TCP_HEADER, priority);
2037 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2041 /* Reserve space for headers and prepare control bits. */
2042 skb_reserve(skb, MAX_TCP_HEADER);
2044 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2045 TCP_SKB_CB(skb)->sacked = 0;
2046 skb_shinfo(skb)->gso_segs = 1;
2047 skb_shinfo(skb)->gso_size = 0;
2048 skb_shinfo(skb)->gso_type = 0;
2051 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
2052 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2053 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2054 if (tcp_transmit_skb(sk, skb, 0, priority))
2055 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2058 /* WARNING: This routine must only be called when we have already sent
2059 * a SYN packet that crossed the incoming SYN that caused this routine
2060 * to get called. If this assumption fails then the initial rcv_wnd
2061 * and rcv_wscale values will not be correct.
2063 int tcp_send_synack(struct sock *sk)
2065 struct sk_buff* skb;
2067 skb = tcp_write_queue_head(sk);
2068 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
2069 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2072 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
2073 if (skb_cloned(skb)) {
2074 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2077 tcp_unlink_write_queue(skb, sk);
2078 skb_header_release(nskb);
2079 __tcp_add_write_queue_head(sk, nskb);
2080 sk_stream_free_skb(sk, skb);
2081 sk_charge_skb(sk, nskb);
2085 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2086 TCP_ECN_send_synack(tcp_sk(sk), skb);
2088 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2089 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2093 * Prepare a SYN-ACK.
2095 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2096 struct request_sock *req)
2098 struct inet_request_sock *ireq = inet_rsk(req);
2099 struct tcp_sock *tp = tcp_sk(sk);
2101 int tcp_header_size;
2102 struct sk_buff *skb;
2103 #ifdef CONFIG_TCP_MD5SIG
2104 struct tcp_md5sig_key *md5;
2105 __u8 *md5_hash_location;
2108 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2112 /* Reserve space for headers. */
2113 skb_reserve(skb, MAX_TCP_HEADER);
2115 skb->dst = dst_clone(dst);
2117 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
2118 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2119 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
2120 /* SACK_PERM is in the place of NOP NOP of TS */
2121 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
2123 #ifdef CONFIG_TCP_MD5SIG
2124 /* Are we doing MD5 on this segment? If so - make room for it */
2125 md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
2127 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
2129 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
2131 memset(th, 0, sizeof(struct tcphdr));
2134 TCP_ECN_make_synack(req, th);
2135 th->source = inet_sk(sk)->sport;
2136 th->dest = ireq->rmt_port;
2137 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
2138 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2139 TCP_SKB_CB(skb)->sacked = 0;
2140 skb_shinfo(skb)->gso_segs = 1;
2141 skb_shinfo(skb)->gso_size = 0;
2142 skb_shinfo(skb)->gso_type = 0;
2143 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2144 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2145 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2147 /* Set this up on the first call only */
2148 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2149 /* tcp_full_space because it is guaranteed to be the first packet */
2150 tcp_select_initial_window(tcp_full_space(sk),
2151 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2156 ireq->rcv_wscale = rcv_wscale;
2159 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2160 th->window = htons(min(req->rcv_wnd, 65535U));
2162 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2163 tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
2164 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
2165 TCP_SKB_CB(skb)->when,
2168 #ifdef CONFIG_TCP_MD5SIG
2169 md5 ? &md5_hash_location :
2175 th->doff = (tcp_header_size >> 2);
2176 TCP_INC_STATS(TCP_MIB_OUTSEGS);
2178 #ifdef CONFIG_TCP_MD5SIG
2179 /* Okay, we have all we need - do the md5 hash if needed */
2181 tp->af_specific->calc_md5_hash(md5_hash_location,
2184 skb->h.th, sk->sk_protocol,
2193 * Do all connect socket setups that can be done AF independent.
2195 static void tcp_connect_init(struct sock *sk)
2197 struct dst_entry *dst = __sk_dst_get(sk);
2198 struct tcp_sock *tp = tcp_sk(sk);
2201 /* We'll fix this up when we get a response from the other end.
2202 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2204 tp->tcp_header_len = sizeof(struct tcphdr) +
2205 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2207 #ifdef CONFIG_TCP_MD5SIG
2208 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2209 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2212 /* If user gave his TCP_MAXSEG, record it to clamp */
2213 if (tp->rx_opt.user_mss)
2214 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2217 tcp_sync_mss(sk, dst_mtu(dst));
2219 if (!tp->window_clamp)
2220 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2221 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2222 tcp_initialize_rcv_mss(sk);
2224 tcp_select_initial_window(tcp_full_space(sk),
2225 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2228 sysctl_tcp_window_scaling,
2231 tp->rx_opt.rcv_wscale = rcv_wscale;
2232 tp->rcv_ssthresh = tp->rcv_wnd;
2235 sock_reset_flag(sk, SOCK_DONE);
2237 tcp_init_wl(tp, tp->write_seq, 0);
2238 tp->snd_una = tp->write_seq;
2239 tp->snd_sml = tp->write_seq;
2244 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2245 inet_csk(sk)->icsk_retransmits = 0;
2246 tcp_clear_retrans(tp);
2250 * Build a SYN and send it off.
2252 int tcp_connect(struct sock *sk)
2254 struct tcp_sock *tp = tcp_sk(sk);
2255 struct sk_buff *buff;
2257 tcp_connect_init(sk);
2259 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2260 if (unlikely(buff == NULL))
2263 /* Reserve space for headers. */
2264 skb_reserve(buff, MAX_TCP_HEADER);
2266 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
2267 TCP_ECN_send_syn(sk, tp, buff);
2268 TCP_SKB_CB(buff)->sacked = 0;
2269 skb_shinfo(buff)->gso_segs = 1;
2270 skb_shinfo(buff)->gso_size = 0;
2271 skb_shinfo(buff)->gso_type = 0;
2273 tp->snd_nxt = tp->write_seq;
2274 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2275 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
2278 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2279 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2280 skb_header_release(buff);
2281 __tcp_add_write_queue_tail(sk, buff);
2282 sk_charge_skb(sk, buff);
2283 tp->packets_out += tcp_skb_pcount(buff);
2284 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2286 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2287 * in order to make this packet get counted in tcpOutSegs.
2289 tp->snd_nxt = tp->write_seq;
2290 tp->pushed_seq = tp->write_seq;
2291 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2293 /* Timer for repeating the SYN until an answer. */
2294 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2295 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2299 /* Send out a delayed ack, the caller does the policy checking
2300 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2303 void tcp_send_delayed_ack(struct sock *sk)
2305 struct inet_connection_sock *icsk = inet_csk(sk);
2306 int ato = icsk->icsk_ack.ato;
2307 unsigned long timeout;
2309 if (ato > TCP_DELACK_MIN) {
2310 const struct tcp_sock *tp = tcp_sk(sk);
2313 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2314 max_ato = TCP_DELACK_MAX;
2316 /* Slow path, intersegment interval is "high". */
2318 /* If some rtt estimate is known, use it to bound delayed ack.
2319 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2323 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2329 ato = min(ato, max_ato);
2332 /* Stay within the limit we were given */
2333 timeout = jiffies + ato;
2335 /* Use new timeout only if there wasn't a older one earlier. */
2336 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2337 /* If delack timer was blocked or is about to expire,
2340 if (icsk->icsk_ack.blocked ||
2341 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2346 if (!time_before(timeout, icsk->icsk_ack.timeout))
2347 timeout = icsk->icsk_ack.timeout;
2349 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2350 icsk->icsk_ack.timeout = timeout;
2351 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2354 /* This routine sends an ack and also updates the window. */
2355 void tcp_send_ack(struct sock *sk)
2357 /* If we have been reset, we may not send again. */
2358 if (sk->sk_state != TCP_CLOSE) {
2359 struct tcp_sock *tp = tcp_sk(sk);
2360 struct sk_buff *buff;
2362 /* We are not putting this on the write queue, so
2363 * tcp_transmit_skb() will set the ownership to this
2366 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2368 inet_csk_schedule_ack(sk);
2369 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2370 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2371 TCP_DELACK_MAX, TCP_RTO_MAX);
2375 /* Reserve space for headers and prepare control bits. */
2376 skb_reserve(buff, MAX_TCP_HEADER);
2378 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2379 TCP_SKB_CB(buff)->sacked = 0;
2380 skb_shinfo(buff)->gso_segs = 1;
2381 skb_shinfo(buff)->gso_size = 0;
2382 skb_shinfo(buff)->gso_type = 0;
2384 /* Send it off, this clears delayed acks for us. */
2385 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
2386 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2387 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2391 /* This routine sends a packet with an out of date sequence
2392 * number. It assumes the other end will try to ack it.
2394 * Question: what should we make while urgent mode?
2395 * 4.4BSD forces sending single byte of data. We cannot send
2396 * out of window data, because we have SND.NXT==SND.MAX...
2398 * Current solution: to send TWO zero-length segments in urgent mode:
2399 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2400 * out-of-date with SND.UNA-1 to probe window.
2402 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2404 struct tcp_sock *tp = tcp_sk(sk);
2405 struct sk_buff *skb;
2407 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2408 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2412 /* Reserve space for headers and set control bits. */
2413 skb_reserve(skb, MAX_TCP_HEADER);
2415 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2416 TCP_SKB_CB(skb)->sacked = urgent;
2417 skb_shinfo(skb)->gso_segs = 1;
2418 skb_shinfo(skb)->gso_size = 0;
2419 skb_shinfo(skb)->gso_type = 0;
2421 /* Use a previous sequence. This should cause the other
2422 * end to send an ack. Don't queue or clone SKB, just
2425 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2426 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2427 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2428 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2431 int tcp_write_wakeup(struct sock *sk)
2433 if (sk->sk_state != TCP_CLOSE) {
2434 struct tcp_sock *tp = tcp_sk(sk);
2435 struct sk_buff *skb;
2437 if ((skb = tcp_send_head(sk)) != NULL &&
2438 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2440 unsigned int mss = tcp_current_mss(sk, 0);
2441 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2443 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2444 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2446 /* We are probing the opening of a window
2447 * but the window size is != 0
2448 * must have been a result SWS avoidance ( sender )
2450 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2452 seg_size = min(seg_size, mss);
2453 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2454 if (tcp_fragment(sk, skb, seg_size, mss))
2456 } else if (!tcp_skb_pcount(skb))
2457 tcp_set_skb_tso_segs(sk, skb, mss);
2459 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2460 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2461 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2463 update_send_head(sk, tp, skb);
2468 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2469 tcp_xmit_probe_skb(sk, TCPCB_URG);
2470 return tcp_xmit_probe_skb(sk, 0);
2476 /* A window probe timeout has occurred. If window is not closed send
2477 * a partial packet else a zero probe.
2479 void tcp_send_probe0(struct sock *sk)
2481 struct inet_connection_sock *icsk = inet_csk(sk);
2482 struct tcp_sock *tp = tcp_sk(sk);
2485 err = tcp_write_wakeup(sk);
2487 if (tp->packets_out || !tcp_send_head(sk)) {
2488 /* Cancel probe timer, if it is not required. */
2489 icsk->icsk_probes_out = 0;
2490 icsk->icsk_backoff = 0;
2495 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2496 icsk->icsk_backoff++;
2497 icsk->icsk_probes_out++;
2498 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2499 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2502 /* If packet was not sent due to local congestion,
2503 * do not backoff and do not remember icsk_probes_out.
2504 * Let local senders to fight for local resources.
2506 * Use accumulated backoff yet.
2508 if (!icsk->icsk_probes_out)
2509 icsk->icsk_probes_out = 1;
2510 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2511 min(icsk->icsk_rto << icsk->icsk_backoff,
2512 TCP_RESOURCE_PROBE_INTERVAL),
2517 EXPORT_SYMBOL(tcp_connect);
2518 EXPORT_SYMBOL(tcp_make_synack);
2519 EXPORT_SYMBOL(tcp_simple_retransmit);
2520 EXPORT_SYMBOL(tcp_sync_mss);
2521 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
2522 EXPORT_SYMBOL(tcp_mtup_init);